FDD-YOLO: A Lightweight Multi-scale Prohibited Items Detection Model

X-ray security inspection faces challenges such as severe occlusion, scale variation, and complex background when detecting prohibited items, requiring real-time and accurate detection. Although the YOLO series of models has high inference efficiency, they suffer from problems such as feature redundancy, insufficient fine-grained feature extraction, and limited adaptability to overlapping objects. To overcome these limitations, we propose FDD-YOLO and design three novel modules: (1) The Frequency Domain Decomposition Network (FDDN) in the backbone network enhances the edges of metal objects and the contours of liquid containers by decomposing high-frequency and low-frequency features while reducing computational redundancy; (2) The Deformable Elastic Fusion Pyramid (DEFP) in the neck network adopts dynamic channel allocation and multi-scale deformable convolution to handle the geometric changes of folded and overlapping objects; (3) The lightweight Dual-channel Convolution (DualConv) improves multi-scale feature capture through grouping and point-by-point convolution, thereby reducing the number of parameters while improving the accuracy of small object detection. Tests on the SIXray, HIXray, and private GIX datasets show that FDD-YOLO achieves 2.6%, 3.2%, and 8.6% higher mAP than YOLOv11n, respectively, achieving accuracies of 94.8%, 84%, and 71.8%, respectively. This framework also reduces the number of parameters by 30.6% and the number of FLOPs by 26.9%, achieving an optimal balance between accuracy and efficiency, setting a new technical benchmark for real-time security inspections.